Skip to main content

Automating AI-BOM (CycloneDX 1.5) Generation in CI/CD Pipelines

A Software Bill of Materials (SBOM) is a fundamental artifact in traditional cybersecurity, mapping npm or PyPI dependencies to track CVEs. However, traditional SBOMs are utterly blind to the components that actually define a Machine Learning system: the training datasets, the model weights, the neural architecture, and the cryptographic provenance.

To address this, OWASP released CycloneDX v1.5, introducing native support for the machine-learning-model and data component types. Generating an AI-BOM is no longer just a best practice—it is a strict regulatory requirement under Article 11 (Technical Documentation) of the EU AI Act.

The Anatomy of an AI-BOM​

An effective AI-BOM must capture metadata that goes far beyond a package version. A CycloneDX 1.5 AI-BOM must document:

  1. Component Type: Identifying if the artifact is a machine-learning-model (e.g., a .safetensors file) or data (e.g., a .parquet training set).
  2. Cryptographic Provenance: The SHA-256 hash of the model weights, ensuring protection against "Silent Model Replacement" attacks.
  3. Upstream Resolution: Links to the original vendor or Hugging Face repository.
  4. Licensing: Machine-readable SPDX identifiers (e.g., MIT, Llama-2-Community) to prevent IP contamination.

Manually curating this JSON structure for every model iteration is impossible. It must be automated within the CI/CD pipeline.

Automating AI-BOMs with Veritensor​

The Veritensor CLI acts as a drop-in DevSecOps tool that automatically parses ML artifacts and generates compliant CycloneDX 1.5 AI-BOMs without requiring manual data entry.

Add the following step to your GitHub Actions or GitLab CI pipeline:

# Example GitHub Actions Workflow
steps:
  - name: Checkout Repository
    uses: actions/checkout@v4

  - name: Install Veritensor
    run: pip install veritensor-cli

  - name: Scan Models and Generate AI-BOM
    run: |
      veritensor scan ./models/ \
        --sbom \
        --output-file ai-bom.json
        
  - name: Upload AI-BOM Artifact
    uses: actions/upload-artifact@v3
    with:
      name: CycloneDX-AI-BOM
      path: ai-bom.json

How the Extraction Works: When Veritensor encounters a model file (e.g., model.safetensors), it does not load the massive weight tensors into memory. Instead, it parses the binary header to extract the tensor count, architecture type, and embedded metadata.

Simultaneously, it calculates the SHA-256 hash of the artifact and queries the Hugging Face API to verify the model's exact origin and licensing terms. This data is then formatted into a strict CycloneDX 1.5 JSON schema.

By automating AI-BOM generation, security and legal teams gain immediate, real-time visibility into the AI supply chain, ensuring that restricted licenses (like AGPL) and unverified models never reach production environments.